Repository navigation
Expand file tree
/
Copy pathembedding.go
More file actions
222 lines (212 loc) · 7.62 KB
/
Copy pathembedding.go
File metadata and controls
222 lines (212 loc) · 7.62 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
// Copyright (c) Microsoft. All rights reserved.
package openaiprovider
import (
"bytes"
"context"
"encoding/base64"
"encoding/binary"
"encoding/json"
"fmt"
"maps"
"math"
"time"
"github.com/microsoft/agent-framework-go/embedding"
"github.com/microsoft/agent-framework-go/internal/telemetry"
"github.com/microsoft/agent-framework-go/message"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/packages/param"
)
// EmbeddingGeneratorConfig configures an OpenAI embedding generator.
type EmbeddingGeneratorConfig struct {
// Model is the default embedding model or deployment name.
Model string
// Dimensions is the default vector size. Nil uses the service default.
// Explicit zero is forwarded; supported values depend on the service.
Dimensions *int
}
// NewEmbeddingGenerator creates a text embedding generator using the client's
// Embeddings API. Results follow input order using the service's indexes and contain
// float32 [embedding.VectorOf] vectors, the response model ID, and a local UTC
// creation timestamp for each embedding. Reported prompt and total token usage
// are preserved; omitted counters default to zero and malformed counters return
// an error. Responses must contain one embedding per input with unique, in-range
// indexes. A single-input response may omit its index. Empty batches return an
// error without an HTTP request.
//
// The [embedding.Options.RawRepresentationFactory] may return an
// [openai.EmbeddingNewParams] value or pointer. Nil or other return types use
// default parameters.
// Native model and dimension settings take precedence over request options,
// which take precedence over config. Other AdditionalProperties are not
// interpreted; use native parameters for provider-specific settings.
// The request input is set from the supplied values. The default encoding is
// base64; native parameters may request float vectors instead.
//
// The generator does not own the client and does not require cleanup.
func NewEmbeddingGenerator(client openai.Client, config EmbeddingGeneratorConfig) embedding.Generator[string, *embedding.Of[float32]] {
if config.Dimensions != nil {
config.Dimensions = new(*config.Dimensions)
}
return &embeddingGenerator{client: client, config: config}
}
type embeddingGenerator struct {
client openai.Client
config EmbeddingGeneratorConfig
}
func (g *embeddingGenerator) Generate(ctx context.Context, values []string, options *embedding.Options) (*embedding.GeneratedOf[float32], error) {
if err := ctx.Err(); err != nil {
return nil, err
}
if len(values) == 0 {
return nil, fmt.Errorf("openaiprovider: input batch must not be empty")
}
var params openai.EmbeddingNewParams
if options != nil && options.RawRepresentationFactory != nil {
switch raw := options.RawRepresentationFactory(g).(type) {
case nil:
case openai.EmbeddingNewParams:
params = raw
case *openai.EmbeddingNewParams:
if raw != nil {
params = *raw
}
}
}
model := g.config.Model
dimensions := g.config.Dimensions
if options != nil {
if options.ModelID != nil {
model = *options.ModelID
}
if options.Dimensions != nil {
dimensions = options.Dimensions
}
}
if params.Model == "" {
params.Model = model
if options != nil && options.ModelID != nil && model == "" {
if _, overridden := params.Overrides(); !overridden {
extras := params.ExtraFields()
if _, present := extras["model"]; !present {
extras = maps.Clone(extras)
if extras == nil {
extras = make(map[string]any)
}
extras["model"] = ""
params.SetExtraFields(extras)
}
}
}
}
if param.IsOmitted(params.Dimensions) && dimensions != nil {
params.Dimensions = openai.Int(int64(*dimensions))
}
if params.EncodingFormat == "" {
params.EncodingFormat = openai.EmbeddingNewParamsEncodingFormatBase64
}
params.Input = openai.EmbeddingNewParamsInputUnion{OfArrayOfStrings: values}
telemetry.MarkUsed(telemetry.FeatureOpenAI)
response, err := g.client.Embeddings.New(ctx, params, telemetryRequestOption(""))
if err != nil {
return nil, err
}
if response == nil || !response.JSON.Data.Valid() && response.JSON.Data.Raw() != "" {
return nil, fmt.Errorf("openaiprovider: invalid embedding response data")
}
if len(response.Data) != len(values) {
return nil, fmt.Errorf("openaiprovider: expected %d embeddings, got %d", len(values), len(response.Data))
}
result := &embedding.GeneratedOf[float32]{
Embeddings: make([]*embedding.Of[float32], len(values)),
}
for i, item := range response.Data {
var index int64
if raw := item.JSON.Index.Raw(); raw != "" {
if !item.JSON.Index.Valid() || json.Unmarshal([]byte(raw), &index) != nil {
return nil, fmt.Errorf("openaiprovider: invalid embedding response index at position %d", i)
}
} else if len(values) > 1 {
return nil, fmt.Errorf("openaiprovider: missing embedding response index at position %d", i)
}
if index < 0 || index >= int64(len(values)) {
return nil, fmt.Errorf("openaiprovider: embedding response index %d is out of range", index)
}
if result.Embeddings[index] != nil {
return nil, fmt.Errorf("openaiprovider: duplicate embedding response index %d", index)
}
vector, err := openAIEmbeddingVector(item)
if err != nil {
return nil, fmt.Errorf("openaiprovider: embedding at position %d: %w", i, err)
}
result.Embeddings[index] = &embedding.Of[float32]{
Vector: vector, ModelID: response.Model, CreatedAt: time.Now().UTC(),
}
}
if response.JSON.Usage.Valid() {
inputTokens, err := openAIEmbeddingTokenCount(response.Usage.JSON.PromptTokens.Raw())
if err != nil {
return nil, fmt.Errorf("openaiprovider: invalid prompt token count: %w", err)
}
totalTokens, err := openAIEmbeddingTokenCount(response.Usage.JSON.TotalTokens.Raw())
if err != nil {
return nil, fmt.Errorf("openaiprovider: invalid total token count: %w", err)
}
result.Usage = &message.UsageDetails{
InputTokenCount: inputTokens,
TotalTokenCount: totalTokens,
}
} else if raw := response.JSON.Usage.Raw(); raw != "" && raw != "null" {
return nil, fmt.Errorf("openaiprovider: embedding response has invalid token usage")
}
return result, nil
}
func openAIEmbeddingVector(item openai.Embedding) (embedding.VectorOf[float32], error) {
var vector embedding.VectorOf[float32]
raw := bytes.TrimSpace([]byte(item.JSON.Embedding.Raw()))
if len(raw) == 0 || bytes.Equal(raw, []byte("null")) {
return nil, fmt.Errorf("missing embedding vector")
}
if raw[0] == '"' {
var encoded string
if err := json.Unmarshal(raw, &encoded); err != nil {
return nil, fmt.Errorf("invalid embedding vector: %w", err)
}
data, err := base64.StdEncoding.DecodeString(encoded)
if err != nil {
return nil, fmt.Errorf("invalid base64 embedding vector: %w", err)
}
if len(data)%4 != 0 {
return nil, fmt.Errorf("base64 embedding vector length must be a multiple of 4")
}
vector = make(embedding.VectorOf[float32], len(data)/4)
for i := range vector {
vector[i] = math.Float32frombits(binary.LittleEndian.Uint32(data[i*4:]))
}
} else {
var elements []*float32
if err := json.Unmarshal(raw, &elements); err != nil {
return nil, fmt.Errorf("invalid embedding vector: %w", err)
}
vector = make(embedding.VectorOf[float32], len(elements))
for i, value := range elements {
if value == nil {
return nil, fmt.Errorf("embedding vector contains a null element")
}
vector[i] = *value
}
}
return vector, nil
}
func openAIEmbeddingTokenCount(raw string) (int64, error) {
if raw == "" {
return 0, nil
}
var count *int64
if err := json.Unmarshal([]byte(raw), &count); err != nil {
return 0, err
}
if count == nil {
return 0, fmt.Errorf("token count must be an integer")
}
return *count, nil
}